Extending Input Channel Using Global Feature Image for Convolutional Neural Networks
A. Abduraimjonov, Hojong Choi, Jaepil Ko · 2019
In the deep learning paradigm, neural networks learn geometric variations to obtain the geometric invariance. Before deep learning emerged, such invariance has been dealt with feature engineering. The success of feature engineering techniques largely depends on a combination of local and global features. Global features can be leaned by deep learning but they are usually achieved in deep layers. In this paper, we propose a new simple training strategy using global feature images as the inputs of the convolutional neural networks, in which the images are generated by image processing techniques. To demonstrate the usefulness of our approach, we set up a toy problem on the MNIST dataset and obtain the best accuracy of 99.51% with the proposed method from our experiments.